December 1, 2010

Day 1 - Linux Containers (LXC)

This article was written by Ken Barber (http://bob.sh/).

Introduction

Linux Containers (abbreviated as LXC) provide a kind of virtual machine that operates within an operating system. For clarity and to keep the semantical pedants happy, this kind of virtualisation is often dubbed "Operating system-level virtualisation." This provides an alternative (or arguably a subset) to what is known as "platform virtualisation" solutions such as KVM, Xen or Vmware.

In the past, we have often used chroot jails to achieve this kind of effect, and products like LXC are simply a modern evolution of this old idea by including isolation and resource management through extensions to the kernel.

This is by no means a new idea. Other examples of technologies like this already include:

  • Solaris Zones
  • BSD Jails
  • OpenVZ
  • Linux vserver
  • FreeVPS

What makes LXC the topic of discussion, here, is that its development has recently made it into the Linux kernel as an official mechanism for providing containers. While these features are always improving, LXC is reaching a stage where it can be considered for production (depending on your level of conservatism, of course), and I feel that it deserves some well earned attention.

Platform vs. OS-level virtualisation

LXC, as mentioned, provides operating system-level virtualisation. For better or worse this means programs still run in the main OS, so only one kernel is ever in operation. This may be an advantage or a disadvantage depending on what you are trying to achieve. Having said that, it provides a level of isolation that is similar to other virtualisation techniques.

For example, if you wish to run a large number of Apache instances while ensuring they are isolated from each other, then LXC is a good fit. Doing this same job using platform virtualisation can add resource overhead thereby not providing the same level of scalability.

I often found that using full virtualisation was wasteful in cases where the instance was created for isolation purposes (which can be handy) but the usage itself was very low. Small mysql databases often need isolation due to differing cache and tuning requirements (and my desire to not have an application kill a shared database platform) but having extra copies of net-snmp and ssh and other system and support tools was wasteful.

On the other hand, if you want to provide complete environments to different groups or customers including SSH, specific OS distributions and kernel revisions then platform virtualisation might be a better fit.

Its all about common sense and finding the right fit. Sometimes the best fit is often a hybrid one. The good news is that libvirt supports LXC now, which means provisioning LXC and KVM (for example) can be managed with the same tool. Also, you can run LXC containers inside a platform virtualised environment. This means you can have LXC containers inside KVM for example, allowing you to create quite interesting mixes of the technologies.

The Technology

LXC works to provide instance isolation and resource control by using several techniques:

  • chroot
  • namespace
  • cgroups

A chroot is an old technique for providing application security through file access isolation. It provides a way to "re-root" a program so that it is not able to access files above that directory. For example if you chroot'd an application to /srv/foo, the application would see / (which is really /srv/foo on the host) only and not be able to access /srv/bar.

The namespace feature was developed as part of LXC. It allows you to isolate resources from each other. These resources can be process IDs, networking information (routes, iptables, interfaces etc.), ipc, and mounts. Much like a chroot works for files, namespace control allows you to isolate process lists so container A cannot see container B's processes.

cgroups was also developed as part of LXC. It allows you to logically bundle together resources and apply resource management, accounting, isolation and other policies to them. This feature provides you with proper control of your container, not just by hiding information but by providing constraints around it.

Together, the complete effect is that a virtual environment created with LXC can be created that in effect looks like a real machine (or perhaps a full virtual machine) but really the processes and functions within it run within the host os's kernel.

Installation

This installation guide here is based around Ubuntu or Debian. A lot of the concepts are the same however on other distributions, and the user space tools for LXC should be roughly identical.

To begin with, we'll need some tools. You will need to install the following packages:

apt-get install lxc bridge-utils debootstrap

Now, for networking its best to place your primary interface inside a bridge. You can achieve this by modifying /etc/network/interfaces as such:

# The primary network interface
#allow-hotplug eth0
#iface eth0 inet dhcp

auto br0
iface br0 inet dhcp
       bridge_ports eth0
       bridge_fd 0
       bridge_maxwait 0

To configure cgroups, you need a special mount (its much like the proc or sys filesystem). Create a new entry in /etc/fstab like so:

cgroup          /cgroup         cgroup  defaults        0       0

Then create the /cgroup directory and mount it.

mkdir -p /cgroup
mount /cgroup

The /cgroup mount can be poked by hand (like /proc) to adjust various resource settings. Later we'll talk about how resource management can be controlled using lxc user space tools.

Manipulating Containers

Once you have done the preliminary steps, you can now start creating and working with containers. Try creating a container named 'vm0' by typing the following commands:

mkdir -p /var/lib/lxc/vm0
/usr/lib/lxc/templates/lxc-debian -p /var/lib/lxc/vm0/

The command lxc-debian is in fact one of a few a wrapper scripts that simplifies the container creation process. It performs a number of tasks for you:

  • Downloads a minimal debian root using 'debootstrap' from http://ftp.debian.org/ to /var/lib/lxc/vm0/rootfs
  • Configures inittab, hostname, locale and networking for the container
  • Removes any unnecessary services such as hwclock and umountfs
  • Sets the root password to 'root'
  • Creates a configuration file in /var/lib/lxc/vm0/config

Of course you are always free to create your own script or to modify this one. In most cases if you are looking to provision lots of containers that perform similar functions its recommended that you tailor your own template to your needs.

There are a number of lxc command line tools for manipulating containers. Most of these are fairly self-explanatory by looking at their name. Lets start by getting a list of instances:

# lxc-ls

Then getting state information on the instance you just created:

# lxc-info --name vm0

Once prepared, you can start the container in the background with:

# lxc-start -d --name vm0

Then connect to the console of your container using:

# lxc-console --name vm0

At this point you will be prompted with a Linux console login prompt. Log into the console using root/root and take a look around. You will notice that a 'ps auxw' results in only a small number of processes. Try the same in your main OS, and you will see that you can see these processes as well, however the PIDs are different.

Doing a pstree -pl however, allows you to see the relationship to a container easily enough:

|-lxc-start(22715)---init(22716)---dhclient3(22917)
|                                |---getty(23006)
|                                |---getty(23008)
|                                |---getty(23009)
|                                |---getty(23010)
|                                \---login(23007)---bash(23013)

Notice how the container has its own copy of init? This provides a convenient (and familiar) parent supervisor for all processes running in your container. Modifying the containers /etc/inittab works just like the real thing, too.

To view container processes clearly from your real OS use this command:

# lxc-ps --name vm0

Or, to see the processes for all containers (including a column which tells you which container they are in) try:

# lxc-ps --lxc

Now to exit out of your console, type Ctrl-A then q. You can then shutdown your instance with:

# lxc-stop --name vm0

Finally - if you wish to destroy your vm, simply do this with:

# lxc-destroy --name vm0

This will delete the rootfs and configuration permanently.

Resource Management

Resource management is taken care of by the cgroups kernel feature.

CPU resource allocation can be done a number of ways, by either pinning CPU affinity or by providing shares. For example, to control CPU affinity for our process, you can use the lxc-cgroup command. First of all lets find out what the current setting is:

# lxc-cgroup -n vm0 cpuset.cpus
0-1

This means that vm0 can use CPU 0 or 1. You can adjust this by providing a new value:

# lxc-cgroup -n vm0 cpuset.cpus '0'

Now our container processes will only schedule on CPU 0. The alternative is to adjust shares, thereby defining priority of instances:

# lxc-cgroup -n vm0 cpu.shares '2048'

# lxc-crgoup -n vm1 cpu.shares '1024'

The above command signifies that vm0 will get twice as much priority as vm1.

If you are looking to permanent change these settings for a specific container, modify the containers configuration file: /var/lib/lxc/vm0/config (for container 'vm0')

For example, to add a specific cpu affinity and share to the container add the line:

lxc.cgroup.cpuset.cpus = 0
lxc.cgroup.cpuset.shares = 2048

There are lots of settings regarding resource management. Its recommended to read the cpuset.txt documentation (see references) for futher details.

Other Cool Stuff

For those cases where you want to pause a container, freezing and unfreezing instances can be done simply with:

# lxc-freeze --name vm0
# lxc-unfreeze --name vm0

If you want the power of LXC but don't want the hassle of creating an environment first a temporary container for running a job can be created simply with:

# lxc-execute -n temp1 -- 'top'

This container can benefit from cgroups management and namespace isolation just like a real container so may come in handy for wrapping application executions.

For fun, I recommend trying some of the other templates available in /var/lib/ lxc/templates. If you install the package 'febootstrap' you can try out lxc-fedora which installs a mini Fedora root for you to use inside Debian and Ubuntu.

Conclusion

I have only touched upon a number of functions within this LXC introduction, so I recommend researching further if you wish to implement this in the wild.

In the mass-hysteria that is the 'Cloud' it's easy to forget that the solution is all about the problem, not the other way around. There is no doubt that at times containers and LXC will provide a neat alternative to platform virtualisation but sometimes this will be the other way around. This of course as usual depends on the problem at hand.

As you can see, LXC is pretty straight-forward. Be sure to try alternative container tools I listed as well, as LXC may not suite your needs (or taste) exactly. Lasty, anything new and adventurous should be thoroughly tested before production deployment to avoid late night adventures.

Further reading:

5 comments :

Cor Cornelisse said...

We (www.hyves.nl) are using LXC in a large production environment for over more than 6 months now.

We're running up to 24 containers on a dual processor Westmere architecture running Gentoo Linux.

This gave us the ability to scale some non-multi-threaded software on multi-core-systems.

We did have some stability issues at start, but traced this back to the unionfs setup. Having replaced this with aufs fixed it. And it's been running stable ever since!

Palle said...

Typo, unless you meant to say:

brdige_fd 0

Jordan Sissel said...

@Palle, thanks - fixed! :)

Albert P. Tobey said...

I did some work with LXC about two years ago. If you're trying it out, be sure to test the snot out of NFS interactions if NFS is involved at all.

Because of the way LXC interacts with the Linux VFS, NFS mounts that go stale can hose everything up and not recover even after the server does.

That may be fixed now, though.

Kip said...

I assume after the bridge configuration networking must be restarted?